$Cerebras(CBRS.US) is indeed an impressive company. You can try out the chatbot on its official website; the token generation speed is mind-blowing: for the same question, deepseek flash streams the answer out word by word, while cerebras almost instantly completes the response. Here are two data points from Cerebras: 485 tokens took only 0.41s, and 1472 tokens took only 1.51s.

What does this mean? It means the company's wafer-scale approach can drastically reduce training and inference times. Large model companies can train models faster (reportedly cutting the time to one-fifth of the original), and users will spend less time waiting for results (or not wait at all).

If the company's solution becomes widely adopted, its market cap and stock price could reach for the stars. The problem lies right here: the competition is too strong (in my opinion). NVIDIA has locked down the upstream and downstream supply chain. NVIDIA's GPUs, Google's TPUs, and others occupy most of TSMC's production capacity. How to convince the supply chain to let core value flow through Cerebras and be captured by the company is a key angle to watch in the medium to long term.

Bullish in the short term, cautious in the medium to long term.